autodev

autonomous software-development loop for oh-my-pi

中文版文档 · Architecture · Testing · Design Decisions · ADRs · Installation · Changelog

autodev is a meta-agent that orchestrates LLM subagents through a YAML-persisted state machine to complete coding tasks autonomously. It does not directly modify files or run commands -- it operates a state machine that does. v1.1.0 adds named subagents with tool whitelists enforced by omp frontmatter — no more prompt-only safety boundaries.

Quick Start

# 1. Install
omp plugin install @williamcodebox/autodev

# 2. Restart omp or reload extensions

# 3. Start
/autodev "add user authentication to the API"

For detailed installation options (including migration from v1.0.0 tarball installs), see INSTALL.md.

Architecture at a Glance

autodev runs a five-phase loop, each phase enforced by the state machine:

flowchart LR
    RECON-PLAN --> RECON --> PLAN --> EXECUTE --> FINAL --> DONE
    RECON -.->|low confidence| RECON-PLAN
    EXECUTE -.->|blocked| EXECUTE

Loading

• RECON-PLAN: autodev-scout agent decides what to investigate
• RECON: Isolated scout instances per dimension return structured dossiers with file:line evidence
• PLAN: TwoRoundGate — autodev-gatekeeper R1 draft + R2 fan-out of N isolated adversarial auditors
• EXECUTE: autodev-implementer runs Design→Implement→Verify within .omp/ firewall constraint
• FINAL: Build standard + final check

Full architecture diagram and details >

Key Features

• Named subagents — tool whitelists enforced by omp frontmatter; no prompt-only safety boundaries
• YAML-persisted state machine — crash recovery via resume anchor; all state changes auditable
• TwoRoundGate — adversarial acceptance criteria review (R1 draft, R2 N isolated auditors per constitution)
• Dynamic recon dimensions — LLM decides what to investigate per task, not hardcoded templates
• HITL/HOTL/auto — three modes, same core loop, zero divergence
• Tool-layer hardening — machine gate pass requires verify evidence; HOTL pause is machine-enforced
• Context budget guardrails — hard tool-level gate prevents context overflow
• Slice boundary handoff — prevents context rot across long tasks

Learn More

Document	What's Inside
Architecture	Full loop diagram, all 8 design features, project structure
Testing	Two-axis methodology, coverage matrix, running tests
Design Decisions	Why YAML? Why separate tools? Why hard gates? Limitations
Installation	Global vs project-level install, verification, uninstall

License

MIT